Zhenyu Li

It could be sort of weird. I guess it is a Pytorch bug. Maybe you should try solutions on google.

Yes. But that can be much later, since I'm looking for a PhD position.

Both needed. I set channel numbers decreasingly (like 512, 256, 128, 64) by experience and reduce the numbers by experience (reduce each to a small number).

After cropping, there is no need to resize it back to the raw resolution. The resizing will cause a change in pixel size, leading to performance degradation of models. That's...

Yes. But that can be much later, since I'm looking for a PhD position.

You should set `front_camera_id` as 0 for KITTI. https://github.com/zhyever/SimIPU/blob/5b346e392c161a5e9fdde09b1692656bc7cd3faf/tools/data_converter/kitti_converter.py#L292 :D Since the released codes are only supporting pre-training on KITTI, data preparation is similar to standard mmdet3d. So, you can...

Sorry that I missed your problems since I was busy recently. There is a problem with my last answer. You should set `front_camera_id=2`. Actually, I recommend that you clone the...

Oh, this issue is caused by the key of `box2d_camera` in `dp_sampler`. In 'tools/create_data.py', you can find the calling of `create_groundtruth_database`, which is used to generate the sampled objects for...

Let me explain why there are problems. We first conduct experiments on KITTI dataset, where the used images come from the second camera. So, when creating the KITTI, all PX...

I will have a check from scratch ASAP and update this repo. Btw, that's the problem only for the Moca training (our downstream task on 3D detection). While the gt_sampler...